Network based on SOM (Self-Organizing-Map) modules combined with statistical decision tools

The neural network discussed in this paper is a self trained network for LArge Memory STorage And Retrieval (LAMSTAR) of information. It employs features such as forgetting, interpolation, extrapolation and filtering, to enhance processing and memory efficiency. The network is based on SOM (Self-Organizing-Map) modules which are modified such that only a limited number of neurons per each module is processed at a given iteration. It employs arrays of link-weight vectors to channel information vertically and horizontally through the network. Direct feedback and up/down counting serve to set these link weights as is a higher-hierarchy performance evaluator element which also provides high level interrupts. Pseudo random modulation of the link weights prevents dogmatic network behavior. The input word is a coded vector X of sub-words (sub-vectors) xi. Each sub-word corresponds to a different category (feature, attribute) of the input word and is processed by a different SOM-type module. The authors have applied the network to simulated medical diagnosis. With adequate input coding, other applications are possible, such as scene recognition and speech recognition. Diagnosis is facilitated by the interpolation/extrapolation capabilities above